In this paper, we propose a novel strategy of filter design for filter bank multicarrier (FBMC) based cognitive radio (CR) systems, and the key idea is to directly optimize the filter coefficients with the objective and constraints of sidelobe suppression. Specifically, a slight modification is implemented to an existing nearly perfect reconstruction (NPR) filter to achieve a better sidelobe suppression than that of the original filter, while maintaining the property of the NPR. Conducted numerical results show that: 1) the sidelobe suppression of the filter could be constrained segment by segment by properly setting the constraints of the proposed scheme; 2) much better sidelobe suppression at specified segments could be achieved in FBMC based CR systems.